Neural network based motion segmentation for accelerometer applications

Of several research issues related to motion interaction using inertia measurement units, faster motion segmentation without accuracy loss has recently been raised. Instead of using excessive filtering that produces time delay or tricky use of multiple thresholds that cause difficulty in parameter optimization, this poster demonstrates that time series prediction using neural networks significantly decreases time delay and guarantees rigid motion segmentation by detecting end points in accelerometer signals. According to a general pattern recognition procedure, feature selection is made by a filtering method and the optimal structure is determined by cross validation. Radial basis function networks and Multi-Layer Perceptrons (MLPs) are tested and the results are compared with the conventional methods to evaluate accuracy and time delay in a handwriting case in 3D space. This study confirms that MLP shows the best accuracy and shortens the time delay by 1/4~1/3 compared to the conventional methods.